Image integration method of convolutional neural network based on selective feature connection

Abstract

The invention discloses an image integration method of a convolutional neural network based on selective feature connection. The method comprises the following steps: respectively solving average features of low-level features and high-level features; subtracting the average feature of the low-level feature from the obtained average feature of the high-level feature to obtain a score of the key feature map; scaling the average features of the high-level features; carrying out Softmax normalization processing to obtain a characteristic Z; and performing maximum value normalization processing onthe feature Z to obtain an attention score. According to the image integration method of the convolutional neural network based on selective feature connection, feature map information can be betterintegrated based on a high-low-layer feature fusion mode of selective feature connection, learned features are more effectively utilized, and the parameter amount is not increased. The structure of the convolutional neural network is optimized, the performance of the network is improved, and the method realize great significance in a shallow convolutional neural network, and enables the shallow convolutional neural network to be applied to more fields.

Description

technical field [0001] The invention belongs to the technical field of convolutional neural networks, and in particular relates to an image integration method of a convolutional neural network based on selective feature connections. Background technique [0002] In recent years, the study of network architecture has attracted a lot of attention. Nowadays, many excellent network architectures have been proposed one after another. GoogLeNet builds a 22-layer convolutional neural network, but it reduces the number of parameters from 60 million to 4 million by using the Inception model. VGGNet proves that using a small convolution filter to increase the depth of the network can effectively improve the effect of the model. However, increasing the depth of the network cannot simply be stacked to make the network layer upon layer. Due to the problems of vanishing and exploding gradients, which make deep networks difficult to train, adding more layers to a properly deep model may...

AI Technical Summary

Problems solved by technology

[0005] In convolutional neural networks, the fusion of high-level and low-level features is an effective way to improve network performance. However, low-level features have problems with background confusion and semantic ambiguity. Blurred, resulting in poor network performance

Images